dany/libwebp
1
0
Fork 0
mirror of https://github.com/webmproject/libwebp.git synced 2025-07-16 05:49:51 +02:00
Code Issues Actions Packages Projects Releases Wiki Activity

introduce WebPMemToUint32 and WebPUint32ToMem for memory access

Browse Source 
it uses memcpy() when unaligned memory write is tricky

Change-Id: I5d966ca9d19e9b43ac90140fa487824116982874
This commit is contained in:
Pascal Massimino
2015-12-02 12:09:06 +01:00
parent 010ca3d10d
commit 2c08aac81a
11 changed files with 112 additions and 108 deletions
Download Patch File Download Diff File Expand all files Collapse all files

87
src/dsp/dec_sse2.c
View File

@ -207,10 +207,10 @@ static void Transform(const int16_t* in, uint8_t* dst, int do_two) {
dst3 = _mm_loadl_epi64((__m128i*)(dst + 3 * BPS));
} else {
// Load four bytes/pixels per line.
dst0 = _mm_cvtsi32_si128(*(int*)(dst + 0 * BPS));
dst1 = _mm_cvtsi32_si128(*(int*)(dst + 1 * BPS));
dst2 = _mm_cvtsi32_si128(*(int*)(dst + 2 * BPS));
dst3 = _mm_cvtsi32_si128(*(int*)(dst + 3 * BPS));
dst0 = _mm_cvtsi32_si128(WebPMemToUint32(dst + 0 * BPS));
dst1 = _mm_cvtsi32_si128(WebPMemToUint32(dst + 1 * BPS));
dst2 = _mm_cvtsi32_si128(WebPMemToUint32(dst + 2 * BPS));
dst3 = _mm_cvtsi32_si128(WebPMemToUint32(dst + 3 * BPS));
}
// Convert to 16b.
dst0 = _mm_unpacklo_epi8(dst0, zero);
@ -236,10 +236,10 @@ static void Transform(const int16_t* in, uint8_t* dst, int do_two) {
_mm_storel_epi64((__m128i*)(dst + 3 * BPS), dst3);
} else {
// Store four bytes/pixels per line.
*(int*)(dst + 0 * BPS) = _mm_cvtsi128_si32(dst0);
*(int*)(dst + 1 * BPS) = _mm_cvtsi128_si32(dst1);
*(int*)(dst + 2 * BPS) = _mm_cvtsi128_si32(dst2);
*(int*)(dst + 3 * BPS) = _mm_cvtsi128_si32(dst3);
WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(dst0));
WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(dst1));
WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(dst2));
WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(dst3));
}
}
}
@ -262,10 +262,10 @@ static void TransformAC3(const int16_t* in, uint8_t* dst) {
const __m128i m3 = _mm_subs_epi16(B, d4);
const __m128i zero = _mm_setzero_si128();
// Load the source pixels.
__m128i dst0 = _mm_cvtsi32_si128(*(int*)(dst + 0 * BPS));
__m128i dst1 = _mm_cvtsi32_si128(*(int*)(dst + 1 * BPS));
__m128i dst2 = _mm_cvtsi32_si128(*(int*)(dst + 2 * BPS));
__m128i dst3 = _mm_cvtsi32_si128(*(int*)(dst + 3 * BPS));
__m128i dst0 = _mm_cvtsi32_si128(WebPMemToUint32(dst + 0 * BPS));
__m128i dst1 = _mm_cvtsi32_si128(WebPMemToUint32(dst + 1 * BPS));
__m128i dst2 = _mm_cvtsi32_si128(WebPMemToUint32(dst + 2 * BPS));
__m128i dst3 = _mm_cvtsi32_si128(WebPMemToUint32(dst + 3 * BPS));
// Convert to 16b.
dst0 = _mm_unpacklo_epi8(dst0, zero);
dst1 = _mm_unpacklo_epi8(dst1, zero);
@ -282,10 +282,10 @@ static void TransformAC3(const int16_t* in, uint8_t* dst) {
dst2 = _mm_packus_epi16(dst2, dst2);
dst3 = _mm_packus_epi16(dst3, dst3);
// Store the results.
*(int*)(dst + 0 * BPS) = _mm_cvtsi128_si32(dst0);
*(int*)(dst + 1 * BPS) = _mm_cvtsi128_si32(dst1);
*(int*)(dst + 2 * BPS) = _mm_cvtsi128_si32(dst2);
*(int*)(dst + 3 * BPS) = _mm_cvtsi128_si32(dst3);
WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(dst0));
WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(dst1));
WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(dst2));
WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(dst3));
}
#undef MUL
#endif // USE_TRANSFORM_AC3
@ -517,24 +517,17 @@ static WEBP_INLINE void DoFilter6(__m128i* const p2, __m128i* const p1,
}
}
// memcpy() is the safe way of moving potentially unaligned 32b memory.
static WEBP_INLINE uint32_t MemToUint32(const uint8_t* const ptr) {
uint32_t A;
memcpy(&A, (const int*)ptr, sizeof(A));
return A;
}
// reads 8 rows across a vertical edge.
static WEBP_INLINE void Load8x4(const uint8_t* const b, int stride,
__m128i* const p, __m128i* const q) {
// A0 = 63 62 61 60 23 22 21 20 43 42 41 40 03 02 01 00
// A1 = 73 72 71 70 33 32 31 30 53 52 51 50 13 12 11 10
const __m128i A0 = _mm_set_epi32(
MemToUint32(&b[6 * stride]), MemToUint32(&b[2 * stride]),
MemToUint32(&b[4 * stride]), MemToUint32(&b[0 * stride]));
WebPMemToUint32(&b[6 * stride]), WebPMemToUint32(&b[2 * stride]),
WebPMemToUint32(&b[4 * stride]), WebPMemToUint32(&b[0 * stride]));
const __m128i A1 = _mm_set_epi32(
MemToUint32(&b[7 * stride]), MemToUint32(&b[3 * stride]),
MemToUint32(&b[5 * stride]), MemToUint32(&b[1 * stride]));
WebPMemToUint32(&b[7 * stride]), WebPMemToUint32(&b[3 * stride]),
WebPMemToUint32(&b[5 * stride]), WebPMemToUint32(&b[1 * stride]));
// B0 = 53 43 52 42 51 41 50 40 13 03 12 02 11 01 10 00
// B1 = 73 63 72 62 71 61 70 60 33 23 32 22 31 21 30 20
@ -592,7 +585,7 @@ static WEBP_INLINE void Load16x4(const uint8_t* const r0,
static WEBP_INLINE void Store4x4(__m128i* const x, uint8_t* dst, int stride) {
int i;
for (i = 0; i < 4; ++i, dst += stride) {
*((int32_t*)dst) = _mm_cvtsi128_si32(*x);
WebPUint32ToMem(dst, _mm_cvtsi128_si32(*x));
*x = _mm_srli_si128(*x, 4);
}
}
@ -963,7 +956,7 @@ static void VE4(uint8_t* dst) { // vertical
const uint32_t vals = _mm_cvtsi128_si32(avg);
int i;
for (i = 0; i < 4; ++i) {
*(uint32_t*)(dst + i * BPS) = vals;
WebPUint32ToMem(dst + i * BPS, vals);
}
}
@ -977,10 +970,10 @@ static void LD4(uint8_t* dst) { // Down-Left
const __m128i lsb = _mm_and_si128(_mm_xor_si128(ABCDEFGH, CDEFGHH0), one);
const __m128i avg2 = _mm_subs_epu8(avg1, lsb);
const __m128i abcdefg = _mm_avg_epu8(avg2, BCDEFGH0);
*(uint32_t*)(dst + 0 * BPS) = _mm_cvtsi128_si32( abcdefg );
*(uint32_t*)(dst + 1 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1));
*(uint32_t*)(dst + 2 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2));
*(uint32_t*)(dst + 3 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3));
WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32( abcdefg ));
WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1)));
WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2)));
WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3)));
}
static void VR4(uint8_t* dst) { // Vertical-Right
@ -998,10 +991,10 @@ static void VR4(uint8_t* dst) { // Vertical-Right
const __m128i lsb = _mm_and_si128(_mm_xor_si128(IXABCD, ABCD0), one);
const __m128i avg2 = _mm_subs_epu8(avg1, lsb);
const __m128i efgh = _mm_avg_epu8(avg2, XABCD);
*(uint32_t*)(dst + 0 * BPS) = _mm_cvtsi128_si32( abcd );
*(uint32_t*)(dst + 1 * BPS) = _mm_cvtsi128_si32( efgh );
*(uint32_t*)(dst + 2 * BPS) = _mm_cvtsi128_si32(_mm_slli_si128(abcd, 1));
*(uint32_t*)(dst + 3 * BPS) = _mm_cvtsi128_si32(_mm_slli_si128(efgh, 1));
WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32( abcd ));
WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32( efgh ));
WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_slli_si128(abcd, 1)));
WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_slli_si128(efgh, 1)));
// these two are hard to implement in SSE2, so we keep the C-version:
DST(0, 2) = AVG3(J, I, X);
@ -1023,10 +1016,10 @@ static void VL4(uint8_t* dst) { // Vertical-Left
const __m128i lsb2 = _mm_and_si128(abbc, lsb1);
const __m128i avg4 = _mm_subs_epu8(avg3, lsb2);
const uint32_t extra_out = _mm_cvtsi128_si32(_mm_srli_si128(avg4, 4));
*(uint32_t*)(dst + 0 * BPS) = _mm_cvtsi128_si32( avg1 );
*(uint32_t*)(dst + 1 * BPS) = _mm_cvtsi128_si32( avg4 );
*(uint32_t*)(dst + 2 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(avg1, 1));
*(uint32_t*)(dst + 3 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(avg4, 1));
WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32( avg1 ));
WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32( avg4 ));
WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(avg1, 1)));
WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(avg4, 1)));
// these two are hard to get and irregular
DST(3, 2) = (extra_out >> 0) & 0xff;
@ -1050,10 +1043,10 @@ static void RD4(uint8_t* dst) { // Down-right
const __m128i lsb = _mm_and_si128(_mm_xor_si128(JIXABCD__, LKJIXABCD), one);
const __m128i avg2 = _mm_subs_epu8(avg1, lsb);
const __m128i abcdefg = _mm_avg_epu8(avg2, KJIXABCD_);
*(uint32_t*)(dst + 3 * BPS) = _mm_cvtsi128_si32( abcdefg );
*(uint32_t*)(dst + 2 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1));
*(uint32_t*)(dst + 1 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2));
*(uint32_t*)(dst + 0 * BPS) = _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3));
WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32( abcdefg ));
WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1)));
WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2)));
WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3)));
}
#undef DST
@ -1067,13 +1060,13 @@ static WEBP_INLINE void TrueMotion(uint8_t* dst, int size) {
const __m128i zero = _mm_setzero_si128();
int y;
if (size == 4) {
const __m128i top_values = _mm_cvtsi32_si128(MemToUint32(top));
const __m128i top_values = _mm_cvtsi32_si128(WebPMemToUint32(top));
const __m128i top_base = _mm_unpacklo_epi8(top_values, zero);
for (y = 0; y < 4; ++y, dst += BPS) {
const int val = dst[-1] - top[-1];
const __m128i base = _mm_set1_epi16(val);
const __m128i out = _mm_packus_epi16(_mm_add_epi16(base, top_base), zero);
*(int*)dst = _mm_cvtsi128_si32(out);
WebPUint32ToMem(dst, _mm_cvtsi128_si32(out));
}
} else if (size == 8) {
const __m128i top_values = _mm_loadl_epi64((const __m128i*)top);